Wolf Digest — Thursday, June 4, 2026

BrainCause attacks a confound in visual neuroscience: strong activation for a concept (faces, places) does not prove a region represents it, since responses may track correlated visual or semantic cues. The framework couples a generative model with an image-to-fMRI encoding model to synthesize controlled stimulus sets—concept images, counterfactual edits that remove the target while preserving other content, and correlated distractors—then searches for representations that respond specifically to the target. It recovers known functional localizations and proposes new candidate representations across dozens of concepts, validated on predicted and measured fMRI, and shows a large fraction of activation-only localizations are false positives.

Trust Region On-Policy Distillation (TrOPD) stabilizes on-policy distillation, where teacher supervision on student-generated tokens yields unreliable gradients under large teacher-student distribution mismatch. TrOPD restricts the reverse-KL (K1) objective to a trust region where the teacher is reliable, handles outlier tokens via gradient clipping, masking, or forward-KL, and adds off-policy guidance by having the student continue from teacher prefixes under forward KL to steer exploration toward reliable regions. It consistently beats SoTA OPD baselines including OPD, EOPD, and REOPOLD across mathematical reasoning, code generation, and general-domain benchmarks.

KVarN targets KV-cache quantization for long-horizon reasoning decode, where prior methods—evaluated in prefill-like settings—break down because quantization errors accumulate across autoregressive timesteps, driven mainly by incorrect per-token scales. The calibration-free quantizer applies a Hadamard rotation followed by dual-scaling variance normalization across both axes of the K and V matrices, correcting outlying token-scale errors and sharply reducing error accumulation. At 2-bit precision it sets a new state of the art on generative benchmarks including MATH500, AIME24, and HumanEval, with a vLLM implementation released.

This work moves deep-research-agent evaluation from final-answer accuracy to span-level error localization—pinpointing which parts of a long search/tool-use/synthesis trajectory make an answer unreliable. The authors collect 2,790 real trajectories across two agent frameworks, three backbones, and three benchmarks, segment logs into semantic spans, and annotate harmful error spans via LLM-assisted expert review to build TELBench (1,000 instances). Their DRIFT auditor tracks agent claims, checks support in trajectory evidence, and flags spans with unsupported or conflicting claims, improving span-level localization and first-error accuracy by up to 30 percentage points.

NVIDIA OmniDreams is a real-time generative world model for closed-loop autonomous-vehicle simulation, mid- and post-trained from the Cosmos diffusion model on 21k hours of driving to autoregressively generate action-conditioned sensor video. Unlike reconstruction-based neural simulators bounded by their captured data, it synthesizes unobserved phenomena such as extreme weather and unpredictable agent behavior, conditioning each frame on past frames, simulator state, and the immediate driving action; it runs in closed loop with the Alpamayo 1 policy and AlpaSim orchestrator. A world-action model post-trained from it surpasses the VLA-based Alpamayo 1.5 policy on the Physical AI NuRec dataset using one-fifth the parameters.

This paper studies when visual world models help multimodal LLMs reason: world models render concrete stochastic rollouts of possible futures, MLLMs reason abstractly over goals and rules, but rollouts can be plausible yet task-incorrect. The authors frame controlled concrete reasoning—learning to invoke, verify, and integrate visual simulation—and release two human-verified benchmarks, VRQABench (spatial lookahead) and OpenWorldQA (open-domain physical prediction). Their method, Privileged-Future On-Policy Self-Distillation (PF-OPSD), uses ground-truth future videos only as teacher-side privileged context while the deployable student never sees true futures, beating the baseline by 10.6% and 10.9% and improving robustness to noisy or conflicting rollouts.

This work gives a localized mechanistic account of cross-domain interference in multi-domain RL post-training, where training on one domain (math, code, QA, creative writing) degrades others even when full-model gradients are nearly orthogonal—contradicting catastrophic-forgetting and global-conflict explanations. Single-domain RL produces sparse, small-magnitude edits with weak neuron overlap but shared computation routes; the authors prove under a local perturbation model that later-domain harm concentrates in a low-dimensional shared conflict subspace. A brief Re-Math refresh after Code-Math-QA-CW recovers Math from 57.66 to 66.04 (best average 66.39), and a training-free rollback on a sparse proxy conflict set partially restores it.

This Google-affiliated work proposes a "Sleep" paradigm for continual learning, addressing models' inability to transfer temporal in-context knowledge into long-term parameters. Sleep has two stages: Memory Consolidation via Knowledge Seeding, an upward distillation that distills a smaller self's memories into a larger network—instantiated as a Generalized Distillation combining on-policy distillation with RL-based imitation learning—and Dreaming, a self-improvement phase where RL generates a synthetic-data curriculum to rehearse new knowledge and refine capabilities without human supervision. Experiments on long-horizon, continual-learning, knowledge-incorporation, and few-shot generalization tasks support the value of the sleep stage.

Qwen-Image-Flash reframes few-step distillation of visual generative models away from objective design toward the training recipe. Using Qwen-Image-2.0 as a case study, the authors systematically vary three factors across unified text-to-image generation and instruction-guided image editing—data composition, teacher guidance, and task mixture—surfacing several non-obvious behaviors. The takeaway: effective few-step distillation needs not only well-designed objectives but principled organization of the broader pipeline, with those findings motivating the released Qwen-Image-Flash model.

AutoMedBench is a workflow-aware benchmark for autonomous medical-AI research, scoring not just final outputs but a unified five-stage workflow—Plan, Setup, Validate, Inference, Submit—across segmentation, image enhancement, VQA, report generation, and lesion detection. Tasks are long-horizon, averaging 33 agent turns, under Lite and Standard scaffolding tiers. Across thousands of runs, Validate is the weakest stage and Setup the strongest, indicating agents make pipelines executable but fail to verify reliability; verification and submission failures dominate tagged errors at 37.7% and 38.1% while task-understanding errors are rare at 0.9%, and a single fired error code lowers overall score by 48% on average.

Trump signed a new AI executive order Tuesday, under two weeks after rescinding the prior one. Key provisions: a voluntary review system asking companies to share frontier models with the government 30 days before release (down from the shelved draft's 90 days); no mandatory pre-deployment licensing; and a dedicated AI cybersecurity clearinghouse to coordinate security checks with industry. Separately, MIT Technology Review details Anduril and Meta's military AR headset prototype, which envisions ordering drone strikes via eye-tracking and voice and aims to optimize "the human as a weapons system."

MIRA is a source-aware data-selection framework for LLM mid-training, where data are optimized under a pretraining-style objective at near-pretraining scale but curated toward downstream capabilities from heterogeneous sources—a regime where model-based selectors give only implicit signals and semantic selectors assume fixed rubrics. MIRA makes rubric construction part of selection via self-anchored rubric discovery: it first discovers what to evaluate per source group, then distills those judgments into scalable student scorers for full-corpus filtering. On code-oriented mid-training across 21 sources and 5 source groups, it beats baselines on nine code benchmarks and matches the full-corpus run using only half the tokens.

M^3Eval is presented as the first comprehensive benchmark probing memory—rather than perception or reasoning—in multimodal models on long-form video: what models retain, how faithfully information is preserved, and how robust it is under interference. Grounded in cognitive psychology, its tasks isolate distinct memory dimensions. Evaluating representative multimodal models reveals consistent weaknesses: they fail to keep disentangled representations across parallel video streams, show interference patterns unlike human memory, ground memory more reliably in the spatial than the temporal domain, and exhibit limited symbolic memory.

AutoLab is a benchmark for ultra-long-horizon, closed-loop optimization: 36 expert-curated tasks across system optimization, puzzles, model development, and CUDA kernel optimization, each starting from a correct-but-suboptimal baseline that an agent must improve within a fixed wall-clock budget. Across 17 frontier models, the dominant predictor of success is not initial-attempt quality but persistence in repeatedly benchmarking, editing, and incorporating empirical feedback. Claude-opus-4.6 shows strong long-horizon behavior, while most models terminate prematurely or burn their budget with minimal progress, underscoring time-awareness as a distinct capability gap. The benchmark, harness, and task artifacts are open-sourced.

OpenAI details how Wasmer used Codex, running on GPT-5.5, to build a Node.js runtime for the edge, reporting a 10x-20x acceleration in development that compressed the timeline from months to weeks. The case study positions Codex as the agentic coding driver for systems-level work (a WebAssembly-based runtime), part of OpenAI's push to show frontier coding models delivering large engineering-velocity gains on real infrastructure projects rather than toy tasks.

Stratechery analyzes Nvidia's move into PC silicon: the RTX Spark superchip (N1X), built with Microsoft and shipping this fall in Windows laptops from Dell, HP, ASUS, Lenovo, and MSI. The top configuration pairs up to 20 Arm cores with a 6,144-CUDA-core Blackwell GPU, 128GB LPDDR5X, and ~300 GB/s bandwidth over NVLink C2C, targeting on-device 120B-parameter models and million-token contexts. The piece judges it broadly DGX Spark-class—strong at prefill but slower than an M5 Max at decode and on CPU—and frames it against Intel, AMD, Qualcomm, and Apple alongside Nadella's platform pitch.

Alphabet raised the size of its equity offering to $84.75 billion from $80 billion—Google's first sale of new stock since 2005. The bulk is earmarked for AI infrastructure and compute, underscoring how capital-intensive frontier-scale buildout has become and signaling Alphabet's willingness to dilute equity rather than rely solely on cash flow and debt to fund data-center and accelerator spending.

VSTAT (Visual STAte Tracking) targets a capability missing from existing video evals: continuously tracking entities, states, and events across a full clip. It comprises 834 synthetic and real-world clips with 1,500 questions deliberately unanswerable from any single frame or short segment. State-of-the-art MLLMs score far below humans and only modestly above answer-prior baselines despite strong results on prior video benchmarks. Comparing thinking traces against the video stream localizes the failure: models reason and track correctly in text but fail at visually perceiving the events they must track. Agentic video and coding agents do not resolve the gap.

StreamMA replaces the "generate-then-transfer" paradigm in multi-agent reasoning, where end-to-end latency scales linearly with pipeline depth, by streaming each reasoning step to downstream agents as it is produced, pipelining adjacent agents. Streaming also improves accuracy: because early reasoning steps are more reliable than later ones, downstream agents act on trustworthy early content rather than error-prone late chains. The authors give the first closed-form joint analysis of stream, serial, and single protocols (effectiveness ordering, speedup bound, cost ratio). Across eight benchmarks, Claude Opus 4.6 and GPT-5.4, and Chain/Tree/Graph topologies, StreamMA averages +7.3 pp (max +22.4 pp on HMMT 2026), and exposes a per-agent "step-level scaling law" orthogonal to agent-count scaling.

TRON is an online environment substrate for visual-reasoning RL that replaces static curated image-QA datasets with a controllable generator-verifier program: each rollout samples a fresh latent visual state, renders an image, poses a question, and exactly verifies the answer, yielding an unbounded stream of instances at the curriculum's current difficulty. The suite spans 520 environments across five ability buckets (spatial, mathematical, diagram, pattern/logic, counting) and supports both a single all-bucket model and per-bucket specialists with no extra data collection. The paper also analyzes generation reliability, instance/level diversity, near-duplicates, and base-model pass rate by difficulty. RL post-training consistently improves ten external multimodal benchmarks across Qwen3-VL-4B, Qwen2.5-VL-7B, and MiMo-VL-7B-SFT.

MERIT attacks two costs of scaling instruction tuning to heterogeneous mixtures, gradient interference and bandwidth-heavy synchronization, by training partitions independently and reconciling them once in parameter space. A local quadratic theory in a shared flat basin shows weight merging yields curvature-weighted variance reduction, PCA-aligned conflict splitting maximizes the gain along high-curvature directions, and merging acts as spectral filtering with implicit norm regularization. The pipeline estimates dataset-level gradient conflicts, partitions along top PCA conflict axes, fine-tunes each partition without inter-partition communication, then merges via token-weighted averaging. On Qwen2.5-VL-3B (136 Vision-FLAN tasks) it lifts the 8-benchmark average from 54.3 to 57.0, scales to a 7B model on a 1.6M-example, 176-source mixture matching centralized joint training, and transfers to text-only FLAN.

Echo-Infinity is an autoregressive framework for real-time infinite video generation that replaces handcrafted KV-cache schedules and fixed-ratio compression with a learnable evolving memory. A set of Memory Query vectors, updated via attention and gating as frames are evicted from the local window and optimized end-to-end with the video DiTs, filters, abstracts, and compresses arbitrary-length history at constant compute independent of video length, and also serves as a generation prior. A Unified Relative RoPE Recipe anchors sink frames at id 0 and caps the newest frame at the pretrained maximum temporal RoPE id, closing the train-test extrapolation gap. It sets state-of-the-art on long and short video generation and demonstrates 24-hour (>1.3M-frame) real-time rollouts, a first toward practical infinite generation.

OpenAI published a blueprint for democratic governance of frontier AI, proposing a U.S. federal framework organized around safety, resilience, and national security. The document argues for federal-level governance of frontier models, situating safety and security oversight within national-security policy. As a position paper from a frontier lab, it signals how OpenAI wants U.S. regulation structured rather than detailing binding mechanisms, and lands amid active federal debate over AI policy and executive action.

Latent Space and No Priors taped a live crossover with Microsoft CEO Satya Nadella at Build. Nadella positions Microsoft as a "Frontier Intelligence Platform," arguing a platform must create more value outside itself than it captures: customers build on multi-model harnesses (OpenClaw, Scout), draw on enterprise context layers like Work IQ, and accumulate private evals and traces as a new form of "Token IP." He also fields the awkward enterprise economics—token-spend cuts, layoffs, and a Build-vs-Buy shift that pressures the SaaS model Microsoft itself epitomizes.

Cognitive Revolution interviews Cornell/Google researcher Ali Behrouz on Nested Learning, his framework for continual learning that updates different parts of a model at different frequencies—mirroring human working-to-long-term memory—to adapt to context while preserving core knowledge. He recasts every ML component as associative memory compressing a context flow (calling fixed "architectures" an illusion) and builds expressive optimizers that learn update rules and beat Adam and Muon. His follow-up, "Language Models Need Sleep," adds an offline consolidation pass that distills high-frequency layers into slow ones and trains on self-generated synthetic data; the architectures reportedly recall across 10M-token contexts.

OpenAI's Codex, repositioned from a developer coding agent to a broader knowledge-worker agent, is drawing defectors from Anthropic's Claude Code as OpenAI improves its models for longer-horizon, more complex tasks. Internally the productivity surge has backfired: OpenAI engineers have gone from two-to-three code changes a day to more than ten, and the resulting volume has caused outages in the systems managing the company's codebase, per two people with knowledge of the situation. Each change triggers thousands of parallel test-hours for correctness and security checks.

PaddleOCR-VL-1.6 upgrades the 0.9B document-parsing model by targeting the residual errors of v1.5, which concentrate in under-optimized regions where behavior is unstable, data is sparse, or supervision is unreliable. Rather than indiscriminately growing the corpus, a region-aware data optimization framework identifies these weak regions from the prior model, applies targeted enhancement, and improves supervision reliability, followed by a progressive post-training recipe combining curated data selection and reinforcement learning. The result is a new state-of-the-art 96.33% on OmniDocBench v1.6, competitive with much larger top-tier VLMs while remaining compact, and a reusable staged post-training recipe for the PaddleOCR-VL series.

GRAIL is a fully virtual data-generation pipeline for humanoid loco-manipulation that avoids teleoperation and motion capture by composing 3D assets, simulator-ready scenes, and video-foundation-model priors. Starting from fully specified 3D configurations, where object geometry, camera, metric scale, environment depth, and a robot-proportioned character are known before generation and reused during reconstruction, it recovers metric 4D human-object interaction trajectories with reduced depth ambiguity and morphology mismatch. Recovered motions are retargeted to a humanoid and used to train an object-aware latent adaptor for manipulation and a scene-aware tracker for traversal. From 20,000+ generated sequences, sim-to-real egocentric visual policies on a Unitree G1 reach 84% real-world success on object pick-up and 90% on stair-climbing.

Anthropic expanded its Claude Partner Network (launched March, backed by $100M), reporting 40,000+ firm applications and 10,000+ certified consultants, with large systems-integrator commitments: Accenture training 30,000, Cognizant ~350,000 associates, Deloitte 470,000, KPMG 276,000+, and PwC rolling out Claude Code and Cowork. The new Services Track defines three tiers, Select (10 certified / 2 production deployments / 1 story), Preferred (100 / 15 / 3), and Global Premier (1,000 / 100 across 3+ regions / 15 stories), measured uniformly regardless of firm size. A daily-refreshed Claude Partner Hub shows partners their standing versus requirements and lets customers find qualified firms, connectable to Claude via a new MCP connector, with promotions twice yearly (Jan 1, Jul 1) plus an Oct 1, 2026 review.

This Latent Space episode features Axiom Math CEO Carina Hong on formal-verification-driven AI for mathematics. The framing: Axiom, seven months old, solved all 12 Putnam problems (8/12 within the time limit), versus the exam's typical median of 0-1 points and DeepSeek's reported 103/120. Hong argues coding ability is necessary but not sufficient for AGI, and that the "informal bottleneck" (translating intuitive proofs into machine-checkable Lean) is where progress stalls. Her thesis, "verification is scaling and compounding brilliance," treats formal proofs as both a way to force articulation and to let others build on results, applying Verified Generation in both training and inference rather than relying on informal chains.

A Hugging Face blog post extends DPO past chatbot alignment to structured OCR via the DharmaOCR pipeline. The trick: after SFT, use the model's own degenerate outputs (repetition-loop transcriptions) as the rejected examples in (chosen, rejected) pairs—"preference-guided implicit unlikelihood"—so the DPO stage actively steers away from the failure geometry SFT can introduce. Across five model families, the DPO stage cut text-degeneration rate relative to SFT in every case, averaging 59.4% reduction (peak 87.6%, e.g. Nanonets-OCR2-3B from 1.61% to 0.20%), needing only a scoring model rather than human labels.

NVIDIA Research is presenting three physical-AI papers at CVPR unified by training at scale for generalization. GraspGen-X is billed as the first foundation model for zero-shot grasping, trained on billions of simulated grasps to propose reliable grasp poses for any gripper geometry and unseen object without per-embodiment retraining. LCDrive replaces costly text-based chain-of-thought in autonomous driving with compact latent representations so AV stacks reason fast enough on embedded hardware. NitroGen, built on the Isaac GR00T architecture, is a generalist gameplay-AI model trained across tens of thousands of hours of simulated interaction.

Broadcom reported fiscal Q2 (ended May 3) revenue of $22.187 billion, up 48% year over year and roughly in line with March guidance. AI chip revenue jumped 143% to $10.8 billion, now nearly half of total sales, driven by demand for custom XPU accelerators and AI networking silicon from hyperscaler customers. The figures underscore Broadcom's position as the leading merchant supplier of custom AI ASICs and Ethernet-based AI fabric, a counterweight to Nvidia's merchant-GPU dominance.

Apple is on track to launch its overhauled Siri in September, running in part on Google cloud servers using Nvidia chips, according to people familiar with the matter. Apple intends to keep as much inference as possible on-device (iPhone), but certain components will run in the cloud. The arrangement is notable for a company that has historically minimized external infrastructure dependence, and it signals reliance on Google's models and Nvidia's accelerators to ship a competitive LLM-based assistant after repeated delays.

Treasury Secretary Scott Bessent defended Trump's new AI executive order, which directs Treasury to stand up an AI cybersecurity clearinghouse within 30 days to coordinate vulnerability scanning, validation and patch distribution with industry, but only on a voluntary basis. Sen. Mark Warner (D-Va.) pressed Bessent, arguing a voluntary regime puts the banking system and national security at risk and calling the order a "watered-down" version of a stronger draft. Bessent said the only change from a prior draft was a tighter deadline (90 to 30 days) and that the order balances innovation and safety.

A CSIS and Foundation for Defense of Democracies commission report concludes current U.S. cyber forces are "insufficient" and that standing up an independent Cyber Force would cost an estimated $10 billion to $11 billion and take at least a year. The proposed service would absorb most of CYBERCOM's "service-like" force-generation duties (organize, train, equip), requiring roughly 20,000 active-duty personnel, 3,500 to 5,000 National Guard, and 6,000 civilians. The report, tied to Sen. Gillibrand's FY2027 NDAA amendment, recommends a Public Health Service-style officer model; the cited budget is already distributed across services ($7.7B for cyberspace operations in FY2027, $4.1B to CYBERCOM).

The Audio Interaction Model reframes large audio LMs from offline, single-task systems (streaming ASR or voice chat) into one online model running an always-on perceive-decide-respond loop that listens to speech, environment, and instructions in real time and decides when to respond from stream semantics. Audio-Interaction retains offline task execution while adding online general audio instruction following, realized via SoundFlow, an end-to-end framework with streaming-native data construction, comprehension-aware training, and asynchronous low-latency inference. Supporting assets include StreamAudio-2M (2.6M items, 7 abilities, 28 sub-tasks) and Proactive-Sound-Bench. Across 8 benchmarks it preserves competitive performance on mainstream tasks while unlocking real-time ASR, streaming instruction following, and proactive intervention.

Reve 2 and Ideogram 4.0 launched the same day, both foregrounding controllable layout and typography in image generation—advances the authors attribute to stronger captioning/labeling and explicit code for composition, a capability once argued to be "AGI-hard." Ideogram 4.0 is billed as the best open image model on the LMArena leaderboard. The AINews recap notes both are US efforts and strong releases, but flags that GPT-Image-2 still leads the arena rankings by a wide margin.

NVIDIA introduced physical-AI "agent skills" at CVPR—modular, agent-callable workflow components meant to stitch together the fragmented pipeline of physical-AI research: scene reconstruction, edge-case scenario generation, policy training, and evaluation. They pair with the newly announced Cosmos 3 world foundation model plus NVIDIA's simulation libraries. The first AV-focused skills include Neural Reconstruction, which rebuilds drivable scenes from fleet data and re-renders them from novel sensor viewpoints to synthesize the rare "long-tail" driving situations that are hard to collect on the road. The skills are published on GitHub (NVIDIA/skills).

Perplexity is bringing Personal Computer—its multi-model orchestration agent "Computer"—to Windows, where it can act on local files and native Microsoft apps (Excel, Word, PowerPoint, Outlook) plus the web. Computer dispatches a team of agents per query grounded in cited answers, and with the Comet browser can fill forms, book appointments, and reach 400+ tools (Snowflake, Salesforce, HubSpot); remote handoff lets a task start on phone and finish on the PC. Files run in a sandbox with auditable, reversible actions and pre-action alerts. It rolls out first to paid Max and Enterprise Max users via waitlist, extending last month's Mac release.

Meta is considering charging up to $200 a month for the consumer version of its planned AI agent, currently called Hatch, per The Information, a price that would match top-tier subscriptions from established AI labs. The product could launch with agentic capabilities aimed at knowledge work. The pricing target is aggressive for Meta, whose consumer AI has been free and ad-supported, and signals an attempt to monetize high-capability agents directly rather than via the ad model.

Google Research open-sourced the hydrology modeling framework behind Flood Hub on GitHub—a PyTorch package implementing LSTM-based river-forecast models that ingest geographic features (climate, soils, topography, land cover) and weather forecasts to predict daily river flow. It ships two model versions, including the one now in production; a recent benchmarking study reports the upgraded architecture extends the reliable forecast horizon by six days in gauged basins and one day in ungauged basins. Trained on the open Caravan dataset and tested with the Czech Hydrometeorological Institute, it lets national agencies fine-tune on local data while retaining control of it.

SpaceX set an expected IPO price of $135 per share, selling 555.6 million shares in a securities filing to raise about $75 billion at a $1.77 trillion valuation. That would be more than twice the size of any prior IPO on record. While not an AI company, the raise signals the scale of capital that frontier hardware and infrastructure plays—launch, satellite networks, and compute-adjacent buildout—can now command in public markets.

Nvidia acquired Kumo AI, a five-year-old startup selling predictive AI software to enterprises, for more than $400 million, per a person with knowledge of the deal; the purchase was first surfaced by an Nvidia executive in a LinkedIn post. Kumo's models target relational/tabular enterprise data for predictive tasks. The buy expands Nvidia's roster of models that can be optimized for its hardware and offered to enterprises for further customization, extending the company's push beyond chips into a software-and-models stack.

Stripe, Visa, Mastercard and Coinbase plan to form a consortium to issue a new stablecoin, per people familiar with the matter, aiming to challenge Circle (USDC) and Tether (USDT), which together hold roughly 80% of the market. The grouping pairs the two largest card networks with a leading payments processor and a major crypto exchange, giving the prospective coin unusually broad distribution and acceptance reach. The move reflects incumbents moving to capture stablecoin payment rails rather than cede them.

Lockheed Martin downed a Group 3 one-way attack drone for the first time using a Joint Air-to-Ground Missile (JAGM) fired from its GRIZZLY containerized launcher at Yuma Proving Ground. The June 3 live-fire integrated the Sanctum C-UAS battle manager with Fortem R-40 radars for a full detect-track-engage kill chain, completed in under 45 days. GRIZZLY holds up to eight Hellfire/JAGM missiles in a 10-foot shipping container with toolless reload and wireless links, pitched as low-cost layered point defense for bases and ships; it builds on a March Hellfire vertical-launch test and a $25M April investment in Fortem.

Senior Pentagon officials are pushing AI and autonomy to address contested logistics, designated by CTO Emil Michael as one of six Critical Technology Areas. Speaking at GDIT's Emerge conference, R&E lead Robert Mantz outlined three focus areas: AI planning tools tied to logistics data lakes that give commanders actionable decision support, autonomous delivery (uncrewed boats/aircraft to complicate enemy targeting and reduce risk), and demand reduction via vehicle hybridization and forward additive manufacturing. Officials also flagged small modular nuclear reactors for base energy resilience and stressed data trust and validity as decisive for commander decision advantage.

Cohere Labs open-sourced co/plot, a data-visualization tool aimed at fast, reproducible research plotting. It targets a gap between Matplotlib (full script reruns for small tweaks) and Figma (polished output but no reliable data ingestion, forcing error-prone manual tracing): co/plot offers prebaked-but-customizable styling that stays faithful to the underlying data. It was built and stress-tested while producing Tiny Aya—scaling to the 70+ languages in that evaluation—and used to style the technical report. Cohere released it under an open-science rationale, and its community researchers are already using it.

War on the Rocks revisits Michael Swaine's 2024 essay "How to Stop the United States and China from Sliding into War," interviewing him after recent Trump-Xi talks. The Q&A probes whether U.S. military involvement in Iran changes his assessment of the risk of major U.S.-China armed conflict and whether it deters Beijing, and reexamines the flashpoints he earlier identified as raising escalation risk. The discussion centers on prospects for U.S.-China strategic stability amid shifting geopolitics.

A Lawfare podcast episode, "Pope Leo XIV Takes on Silicon Valley," features guests Christopher Hale and Renee DiResta discussing the Vatican's posture toward the technology industry under Pope Leo XIV. The conversation examines the Church's engagement with Silicon Valley on issues spanning AI and the social and ethical questions raised by powerful technology platforms. (Source routed via Google News; summary drawn from the episode title and listed participants, as no article body was available.)

Luma joined Human After All, Webedia-Elephant's new AI Creator Studio launching in Paris, as a creative partner alongside Google and ElevenLabs. At the launch Luma demoed live talent transformed in real time with Ray3.14—the performer stays while the surrounding world changes—and will run summer sessions where creators produce real work targeting tasks that were previously hard: suit-free performance capture, video-to-video that keeps a character consistent across shots, and full-pipeline world generation. The framing keeps the human maker central, with AI extending range; this is a partnership announcement, not a model release.